[Eeglablist] How does transferring ICA matrixes between same-subject data sets affect further processing?

[Makoto Miyakoshi]

Dear Duncan,

Your question is legitimate, Duncan. Actually a very good point, and
describing the current weakpoint of stationary ICA.

> I am not sure what happens with the parts of the signal that did not go
into ICA – it seems like this is unaccounted for by any independent
component?


I assume that you cut out the noisy part (data rejection), run ICA, and
apply the ICA weight to the unrejected data. If so, yes you are right, the
noise is not accounted by any ICs. The result is that you see the artifact
spread to all ICs.


> Would this mean that if I decided to subtract all IC’s from the data to
which the weight and sphering matrixes were transferred, there would be a
remaining signal consisting of all the activity that lacked in the passive
conditions?


Right, if I understand the situation you explained.


> I am assuming this also means that for example, any neck muscle activity
particular to the active conditions cannot be subtracted from the signal in
that data, for the donor data set did not contain this.


Right, again if my understanding is correct.

Strictly speaking, therefore, when you copy ICA weight matrix you should
know that the data that were not provided to ICA will not be decomposed
(particularly if the portion of data are associated with different types of
cognitive/physical/artifact states than those during ICA-learned
datapoints).


> And does this change if there was some minor neck muscle EMG signal
present in the donor set, compared to major activity in the receiving one?


As long as they come from the same source with the same volume conductance
and (scalp) mixing, it works and amplitude does not matter. I'm not sure if
muscles are that linear though.


> I guess this must at least help doing a better source localization of
such EMG signals, although that procedure in itself may be better left for
another discussion.

EMG are difficult for ICA because its source moves and spreads. This is
different issue though.

When you run amica (a plugin for EEGLAB, does better job than infomax), the
result comes with log likelihood values with which you can evaluate how
good the decomposition is for every datapoint. This is important to
evaluate stationarity of data. Because standard infomax in EEGLAB or
single-model AMICA generates one unmixing matrix, this is only valid when
your data have only one (stationary) state. If you have task-rest-task-rest
data, you may want to run separate ICA on task part and rest part
separately, or perform two-(or more) model AMICA to hope to let them
capture different states of data.

Sorry for writing quick and non-native English limitation.

Makoto



On Thu, Oct 20, 2016 at 8:59 AM, <duncan.huizinga at gmail.com> wrote:

>
>
> Hi all,
>
>
>
> I have been discovering about the wonders of EEG research and all the
> impressive possibilities with EEGLab, but face one challenging question I
> have not been able to answer so far, which likely in part because I am
> still not completely certain about the precise nature of the ICA algorithm.
> (This might also mean the title question is somewhat non-sensical; In that
> case please forgive my ignorance.) Let’s assume I do understand, and please
> correct me whenever my reasoning is wrong:
>
>
>
> I recorded EEG data (64 channel) from a number of subjects in various
> bodily states, including sitting, standing, (fast) walking and cycling. The
> active conditions obviously add a considerable amount of noise to the data,
> in the form of excessive EMG, cable sway, sensor displacement and slow
> drifts as a result of sweat. I am not directly looking into any locomotion
> related brain activity, merely its effects on other cognitive processes.
>
> In order to be able to do succesful source separation and localization, I
> figured that one could run ICA on the recordings from the passive
> conditions – assuming that the neural sources of the signals that are of
> interest to us remain stable – and then use the resulting weight and
> sphering matrixes for the active conditions as well (in a similar fashion
> as I have seen it suggested with high pass filtered data in Makoto’s
> preproc. pipeline).
>
>
>
> This is where I get lost, because I am not sure what happens with the
> parts of the signal that did not go into ICA – it seems like this is
> unaccounted for by any independent component? Would this mean that if I
> decided to subtract all IC’s from the data to which the weight and sphering
> matrixes were transferred, there would be a remaining signal consisting of
> all the activity that lacked in the passive conditions?
>
> I am assuming this also means that for example, any neck muscle activity
> particular to the active conditions cannot be subtracted from the signal in
> that data, for the donor data set did not contain this.
>
> And does this change if there was some minor neck muscle EMG signal
> present in the donor set, compared to major activity in the receiving one?
> I guess this must at least help doing a better source localization of such
> EMG signals, although that procedure in itself may be better left for
> another discussion.
>
>
>
> Thanks a lot in advance!
>
>
>
> Regards,
>
> Duncan Huizinga
>
>
>
>
>
>
>
>
>
> On a side note: Some of you may have noticed browsing through the
> EEGLablist online archives can be a bit of a mess, as some text refuses to
> reflow. I made a very basic Stylish-script
> <https://userstyles.org/styles/133970/eeglablist-simple-clean-reflowed>
> (useful addon for Firefox
> <https://addons.mozilla.org/en-US/firefox/addon/stylish/> and Chrome
> <https://chrome.google.com/webstore/detail/stylish/fjnbnpbmkenffdnngjfgmeleoegfcffe?hl=nl>)
> which should solve that issue, and it changes the font into something nice
> (which is easily changed into your own preferences by the way; just look
> into the code).
>
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.
> ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to
> eeglablist-request at sccn.ucsd.edu
>



-- 
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20161024/8a65e166/attachment.html>